🖌️ Vector Paintings
Keyword: VectorPaintDaal
Recommended Strength: 0.5–1.5 (Best at 1.0) for Flux
Illustrious recommended strengths 1.0 to 1.5 works best at 1.3 somewhat subjective.
🎨 Where Vector Meets Emotion
Vector Paintings is a stylistic LoRA that fuses bold vector shapes with an expressive painterly overlay—resulting in vivid images brimming with color, character, and emotion. Think clean lines meets emotional brushwork, with a sense of motion, light, and whimsy woven into every scene.
This LoRA shines with people, creatures, urban scenes, fantasy beasts, and emotion-rich environments. Whether you're painting a rainbow-haired barista, a whimsical flying whale, or a joyful monster barreling through the woods, this LoRA captures the energy and emotion of the moment with the storytelling power of impressionist brushwork layered over modern vector design.
The result? A stylized world that feels animated yet artistic, structured yet emotional. It’s especially great for conveying joy, wonder, mystery, or nostalgia—ideal for storytellers and illustrators alike.
🧠 How to Use It:
Use
VectorPaintDaalin your prompt to trigger the style.Works best at 1.0, but experiment between 0.5 to 1.5 for more subtle or intense stylization.
Works well with single-subject prompts or narrative settings that emphasize light, motion, or a core feeling.
Avoid ultra-photorealistic or heavily literal prompts—this LoRA thrives on creative expression and storytelling tone.
💬 Add to the Colorful Conversation!
If you like what you create, give it a thumbs up and share your first image on the LoRA page. Your creations bring the model to life, spark new inspiration, and help keep this vibrant community growing.
Let's keep the brush strokes flying. 💫
Description
Took several tries to get this one dialed in. Mostly user error as I figure out what Qwen likes and doesn't. But in the end, I am happy with this one.
FAQ
Comments (14)
Qwen version looks great. Was it trained using the same dataset as the Flux version?
I find training for Qwen quite easy and fast, but one needs fairly high rank (ast least dim 16) compared to Flux.
It was the same dataset but with a bit of tweaking on image normalization. I screwed this one up 3 times before I got things actually set to how I wanted them to be. I've been trying a new trainer (AI-Toolkit) so that didn't help with a double learning curve.
@Daalis thanks for the info. By image normalization, do you mean adjusting the color curve of the images?
@NobodyButMeow Nope, not even that complicated. I had an issue where the memory would explode due to the number of buckets due to size variants. While Qwen responds really well to 1024x1024 I had a mix of sizes and buckets so I had to normalize to 768x768. When I say explode I mean EXPLODE, it would go from 2 to 7 seconds per iteration to 55 after a checkpoint write. I always normalize on color, theme, etc before creating the image batches. I don't edit them but I pick and curate. Hope all that makes sense.
@Daalis Yes, all that makes sense. Thanks for the explanation 🙏
triggers not for me.....
edit : it functions Comfy and me got a battle than it worked /edit
You can see all my settings in the embedded Comfy JSON on the images.
yeah maybe skill issues i don't know thanks anyway 😉
@Daalis think comfy got it's stinky day 4 me it seems to work
@cynic2010 Seeds.... Alas. I ran a few of your prompts through on Civit and took out some of the leading style bits and they did seem to work. So simpler or no style seems to be more important for Z to not override perhaps.
@Daalis Yeah, I mentioned it myself: if something overrides a set trigger, it triggers the override instead of the actual LoRA. For example, yours uses an anime trigger ..like “anime”..., but if you write something like ....“photorealistic” or similar in the prompt, it will prioritize that first and only maybe apply the LoRA afterward.
Yesterday ComfyUI was acting up big time for me anyway, so it ended up being a weird mix of both. THANKS either way!
❤️😉🌞👍Keep going
I'm doing some lora stacking experimentation. Was this lora made with the de-distilled model?
With the straight Z Image Turbo model off of huggingface Tongyi-MAI (Tongyi-MAI).
@Daalis Awesome, thanks!